A few expert queries suffices for sample-efficient rl with resets and linear value approximation